I did linear regression of X (independent variable) to M (Mediator)
then I used survival regression to fit X to Y (dependent variable)
With these questions:
a. HOW to correctly do a mediation analysis from X to Y through M with survival regression?
b. If the Mediation() function is available, why the results are so weird? ie. ACME and ADE are so large and have negative values.
C. if the negative values are fine, how to explain them? As I know, they might be explained as the suppressing effects.
I'm new to mediation analysis and I'm using mediation() with R. My results are very strange and I'm not sure if they are correct. I haven't found a very detailed mediation analysis on survival regression, any discussion is very welcome and if anyone can give me some hints I would appreciate it!
Here is the code:
# Mediator model
mediator_formula
When conducting a mediation analysis with mixed models—using linear regression for the mediator (M) and survival regression for the outcome (Y)—there are several important considerations and potential troubleshooting steps. Here's a step-by-step guide:
Steps in Mediation Analysis with Mixed Models
Troubleshooting Common Issues
Becareful when you are using ratio scale especially for parametric test in any inferetial data analysis. The rule is that the dependent variable must be in ratio scale and must be normally distributed. If the dependent variable is not normally distributed, then adjust the data using either log transformation or other method.
To correctly perform a mediation analysis from X to Y through M using survival regression, you should follow a more specialized approach suited for time-to-event data. Firstly, ensure your mediator model (M ~ X) is accurately specified using linear regression, as you have done. Next, for the dependent variable model, apply a survival regression model (e.g., Cox proportional hazards model) to fit Y using both X and M. You should use a mediation analysis technique that accommodates survival outcomes, such as the mediation package in R. The results from the mediate() function might appear strange due to the nature of your data or model specification issues. Large negative ACME and ADE values can indeed indicate suppression effects, where the mediator M has a different effect direction on Y compared to X. This suppression can occur when M reduces the apparent effect of X on Y, making direct and indirect effects appear contradictory. When interpreting negative values, consider the context and the potential for suppression: it implies that while X may increase Y directly, the mediation through M may result in a decrease in Y, thus creating an overall negative indirect effect. It's crucial to review the theoretical underpinnings of your variables and the assumptions of the mediation model, particularly within the context of survival analysis. I recommend consulting advanced mediation analysis resources or seeking advice from a statistician with expertise in survival analysis to ensure the appropriateness of your model and the interpretation of results.
Adeoye Ojo
Hi Adeoye, thank you for the information, I checked the distributions of the mediator, the scaled mediator is normally distributed. As for the dependent variable, it is the incidence of a disease (binary categorical variables, only 1 and 0). Still not sure, in this case, whether it is required that the status of the event also meets whatever distribution.
Andaç Akçakese
Hi Andaç， thank you for this advice. As I searched online, "In the literature such mediations are often drawn as path models in which X represents the predictor, M the mediator and Y the dependent variable. X -> Y is then called c', X -> M becomes a and M -> Y is called b."
I actually checked the regressions step by step. My ACME doesn't equal a*b, nor does ADE equal c', not even by an order of magnitude. Something must be wrong somewhere, and I think I do need the help of a statistician. Thank you.
Wei Wang kindly use this link
https://www.spss-tutorials.com/spss-mediation-analysis-tutorial/
to see example of mediation analysis you could follow.
Run your model one after the other.I believe you should have answers to some of your questions.
2. IF this? "it is the incidence of a disease (binary categorical variables, only 1 and 0)". I suggest your categorical variable should be labelled as( 1= presence and 0= absent) in variable view. See attached pictures for unreal data I just generate to show you examples of how to run the analysis and assumed output you should have.
3. The major issue here is that the dependent variable, which is binary categorical only varied between "0" and "1" is now qualitative, which is supposed to be measured in quantitative to be able to use parametric test (e.g Linear Regression).
4. In this case would you consider Binary Logistic Regression?
- Badges
- Science topic
- Similar topics
- Names